As the diagnosis of neuropsychiatric disorders relies on the subjective judgement of the physician, there is a need for the development of objective biomarkers. Recent studies have reported that non-invasive investigation of functional connectivity using resting-state fMRI can serve as a biomarker for psychiatric disorders, and this is a promising development. However, many of the currently available methods are not sufficiently accurate for diagnosis, suggesting limitations of conventional analysis methods. In this project, we will improve the accuracy of resting-state fMRI analysis in mice by analysing time series of brain state information at rest using a linear coupled model, separating brain state information and task-related information during task performance using a linear coupled model, and using optogenetics to identify causal relationships between neuronal responses reflecting brain state information and task-related information Develop methods. Human fMRI experts (Dr Chikazoe from Shintani) and mouse experts in calcium imaging and optogenetics (Dr Kozuma from the Institute of Physiology) are collaborating on the human and mouse studies.

The influence of emotions on human behaviour is one of the most fundamental topics in human science. However, as the subjective emotions of individuals are invisible, it is not easy to construct mathematical models of human behaviour with emotions as variables. In recent years, dramatic advances in machine learning technology have made it possible to decode hidden emotional information from neural and physical information. We are committed to providing new models, concepts and theories in the field of human sciences based on the emotional information decoded from brain activity.

Humans and computers can derive subjective values from sensory events, but such transformation processes are an unexplored area. This study aims to elucidate unknown neural mechanisms by comparing convolutional neural networks (CNNs) with their human counterparts.
Specifically, we optimise CNNs to predict the aesthetic evaluation of paintings and focus on the relationship between CNN representations and brain activity through multi-voxel pattern analysis. Activity in primary visual cortex and higher association cortex was similar to that computed in the shallow and deep CNN layers, respectively. Visual-to-value conversion proved to be a hierarchical process, consistent with a principal gradient linking unimodal to transmodal brain regions (default mode network). In addition, activity in the frontal and parietal lobes was approximated by a goal-driven CNN. As a result, the representation of the hidden layers of the CNN could be understood and visualised through correspondences with brain activity, confirming the similarities between artificial intelligence and neuroscience.
A preprint of this study is available on biorxiv.